AI & MACHINE LEARNING PORTFOLIO

AI Projects by Piyush Landge

A comprehensive catalog of production AI applications, Retrieval-Augmented Generation (RAG) pipelines, deep learning diagnostic toolkits, and machine learning systems engineered by Piyush Landge in Pune, Maharashtra.

Project 01 • Generative AI & LLM Application

AI Flyer Designer

Multi-agent generative design pipeline combining Gemini 2.5 Flash for compositional layout generation and Imagen 3 for 2K high-resolution visual generation.

Problem: SMBs and creators spend hours creating localized marketing flyers with expensive design suites.

Implementation: Dual-model pipeline, structured JSON output enforcement, and React + Tailwind UI canvas.

Verified Results: Sub-8-second flyer generation with full font hierarchy and exportable high-res visual assets.

View Source Code →

Project 02 • Enterprise Search & RAG Architecture

Contextual Retrieval Augmented Generation (RAG)

Production-grade RAG pipeline integrating Qdrant vector database, hybrid dense-sparse retrieval, and Gemini text embeddings with cross-encoder semantic reranking.

Problem: Traditional keyword search fails on semantic nuances, while naive vector search retrieves out-of-context chunks.

Implementation: Contextual chunk prepending, Reciprocal Rank Fusion (RRF), and streaming LLM synthesis.

Verified Results: Reduced hallucination rates in technical document QA benchmarks with accurate citation grounding.

View Source Code →

Project 03 • Model Interpretability & PyTorch Tooling

Deep Learning Auditor & Interpretability Toolkit

Neural network diagnostic toolkit for PyTorch providing Grad-CAM saliency heatmaps, layer activation telemetry, and vanishing gradient detection.

Problem: Deep neural networks operate as black boxes, making edge-case misclassifications difficult to diagnose.

Implementation: PyTorch forward/backward hooks, Grad-CAM saliency mapping, and interactive Streamlit inspection dashboard.

Verified Results: Enabled rapid visual model audits, pinpointing spurious feature correlations and gradient starvation.

View Source Code →

Project 04 • Autonomous Agents & Web Systems

Portfolio & Autonomous AI Agent Interface

Ultra-fast portfolio website integrated with a serverless Gemini 2.5 Flash conversational agent with strict domain knowledge grounding and rate limiting.

Problem: Standard static portfolios are passive and fail to showcase interactive AI capabilities in real time.

Implementation: Vercel serverless edge function, streaming Gemini 2.5 API with system prompt grounding, and lightweight React client.

Verified Results: 100/100 Lighthouse performance baseline with sub-400ms first-token generation response.

Live System Demo →